NEWS IN FOCUS
A RTI FI C I A L I N T E L L I G E NCE
AI firms lure
academics
Shift to industry sparks excitement — and some concern.
BY ELIZABETH GIBNEY
W
hen Andrew Ng joined Google
from Stanford University in
2011, he was among a trickle
of artificial-intelligence (AI) experts in
academia taking up roles in industry.
Five years later, demand for expertise in
AI is booming — and a torrent of researchers is following Ng’s lead. The laboratories
of tech titans Google, Microsoft, Facebook,
IBM and Baidu (China’s web-services giant)
are stuffed with ex-university scientists,
drawn to private firms’ superior computing resources and salaries. “Some people
in academia blame me for starting part of
this,” says Ng, who in 2014 moved again to
become chief scientist at Baidu, working at
the company’s research lab in California’s
Silicon Valley.
Many scientists say that the intense
corporate interest is a boon to AI — bringing vast engineering resources to the field,
demonstrating its real-world relevance
and attracting eager students. But some are
concerned about the more subtle impacts of
the industrial migration, which leaves universities temporarily devoid of top talent,
and could ultimately sway the field towards
commercial endeavours at the expense of
fundamental research.
.
d
e
v
r
e
s
e
r
s
t
h
g
i
r
l
l
A
.
d
e
t
i
m
i
L
s
r
e
h
s
i
l
b
u
P
n
a
l
l
i
m
c
a
M
6
1
0
2
©
4 2 2 | NAT U R E | VO L 5 3 2 | 2 8 A P R I L 2 0 1 6
Robin Li, head of China’s web giant Baidu, unveils the firm’s intelligent digital assistant, Duer.
Private firms are investing heavily in
AI — and in particular in an AI technique
called deep learning — because of its promise
to glean understanding from huge amounts
of data. Sophisticated AI systems might be
able to create effective digital personal assistants, control self-driving cars, or take on a
host of other tasks that are too complex for
conventional programming. And corporate
labs’ resources allow progress that might not
be possible in academic departments, says
Geoffrey Hinton, a deep-learning pioneer
at the University of Toronto in Canada who
took up a post at Google in 2013. The fields of
speech and image recognition, for instance,
had been held up for years by a lack of data to
use in training algorithms and a shortage of
hardware, he says — bottlenecks that he was
able to get past at Google.
“AI is so hot right now. There are so many
opportunities and so few people to work on
them,” says Ng, who says he was attracted by
Google’s reams of data and computing power,
and its ability to tackle real-world problems.
Another temptation is that private firms can
offer “astronomical” wages, says Tara Sinclair,
chief economist at Indeed, a firm headquartered
in Austin, Texas, that aggregates online job listings and has chronicled a rising demand for jobs
in AI in Britain and the United States.
The excitement shows that AI is at a point at
CHINAFOTOPRESS/GETTY
spotted in Kentucky in 2011, but vigorous
surveillance helped to stop it spreading in the
United States.
In South America, the disease tends to take
hold in hot and humid spells. Such conditions
are present in Bangladesh, and the disease
could migrate across south and southeast Asia,
say plant pathologists. In particular, it could
spread over the Indo-Gangetic Plain through
Bangladesh, northern India and eastern Pakistan, warn scientists at the Bangladesh Agricultural Research Institute (BARI) in Nashipur.
Bangladeshi officials are burning government-owned wheat fields to contain the
fungus, and telling farmers not to sow seeds
from infected plots. The BARI hopes to identify
wheat varieties that are more tolerant of the fungus and agricultural practices that can keep it at
bay, such as crop rotation and seed treatment.
It is unknown how wheat blast got to Bangladesh. One possibility is that a wheat-infecting
strain was brought in from South America, says
Nick Talbot, a plant pathologist at the University of Exeter, UK. Another is that an M. oryzae
strain that infects south Asian grasses somehow
jumped to wheat, perhaps triggered by an environmental shift: that is what happened in Kentucky, when a rye-grass strain infected wheat.
To tackle the question, this month Kamoun’s
lab sequenced a fungus sample from Bangladesh. The strain seems to be related to those that
infect wheat in South America, says Kamoun,
but data from other wheat-infecting strains and
strains that plague other grasses are needed to
pinpoint the outbreak’s origins conclusively.
The Open Wheat Blast website might help.
Kamoun has uploaded the Bangladeshi data,
and Talbot has deposited M. oryzae sequences
from wheat in Brazil. Talbot hopes that widely
accessible genome data could help to combat the
outbreak. Researchers could use them to screen
seeds for infection or identify wild grasses that
can transmit the fungus to wheat fields.
Rapid data sharing is becoming more
common in health emergencies, such as the outbreak of Zika virus in the Americas. Kamoun
and Talbot say that their field should follow
suit. “The plant-pathology community has a
responsibility to allow data to be used to combat
diseases that are happening now, and not worry
too much about whether they may or may not
get a Nature paper out of it,” says Talbot.
Last month, Valent’s team reported the first
gene variant known to confer wheat-blast resistance (C. D. Cruz et al. Crop Sci. http://doi.org/
bfk7; 2016), and field trials of crops that bear
the resistance gene variant have begun in South
America. But plant pathologists say that finding
one variant is not enough: wheat strains must be
bred with multiple genes for resistance, to stop
M. oryzae quickly overcoming their defences.
The work could help in the Asian crisis,
says Talbot. “What I would hope for out of this
sorry situation,” he says, “is that there will be
a bigger international effort to identify resistance genes.” ■
IN FOCUS NEWS
GOOGLE DEEPMIND’S TALENT GRAB
Google DeepMind, an artificial-intelligence firm in London, has embarked on a hiring spree since 2014. Its staff declined
to discuss the migration of AI talent, but data gathered by Nature suggest that the firm’s current roster includes at least
144 researchers — almost two-thirds of them drawn from universities.
Researcher origin
Corporate
Academic
160
140
35%
120
144
Google DeepMind
researchers
Massachusetts Institute
of Technology
Cumulative growth in Google DeepMind research staff.
65%
Google
University of Toronto
Total researchers
Around 65% of Google DeepMind’s
research hires came direct from
academia (including joint
appointments, post-PhD and
MSc hires). Many are from
University College London,
where co-founders Demis
Hassabis and Shane Legg
both worked before
starting the firm in 2010.
Google acquires
DeepMind in
January 2014.
100
80
60
40
Imperial College London
University of Montreal
20
University of Cambridge
University College London
University of Alberta
0
2010
University of Oxford
2011
2012
2013
2014
2015
2016
Data collated by Nature from online sources including Scopus, LinkedIn, Google Scholar and personal webpages. ‘Researchers’ excludes most software engineers and developers, and all administrative or other staff. Researchers were identified by title (such as
‘research scientist’ or ‘research engineer’) or previous role. Not all institutions shown in breakdown.
which it can achieve real-world impact — and
companies are the natural way to make this
happen, says Pieter Abbeel, a specialist in AI
and deep learning at the University of California, Berkeley. In the 1950s, a similar job migration occurred in semiconductor research,
when many of the field’s leading figures
were poached to become heads of industrial
research-and-development labs, says Robert
Tijssen, a social scientist at Leiden University
in the Netherlands. Moving academics bring
expertise while extending their new-found
corporate networks back to their former colleagues and students — making the scenario a
“classic win–win situation”, Tijssen says.
CORPORATE COLLABORATIONS
Herman Herman, director of the US National
Robotics Engineering Center based at Carnegie
Mellon University in Pittsburgh, Pennsylvania,
subscribes to that view. In 2015, car-hailing app
Uber, which was collaborating with the centre,
hired almost 40 of his 150 researchers, mainly
those working on self-driving cars. Reports at
the time suggested that the centre was left in
crisis, but Herman says this was overblown;
the project was one of dozens across Carnegie
Mellon’s Robotics Institute, which has about
500 faculty members. The move was a chance
to bring in new blood, and shortly afterwards,
Uber donated US$5.5 million to support student and faculty fellowships at the institute.
Meanwhile, the publicity raised the profile of
the centre’s work, says Herman — and student
applications are up.
The loss of expertise in academia concerns
Yoshua Bengio, a computer scientist at the
University of Montreal in Canada, which has
also seen a surge in graduate-student applications. If industry-hired faculty members do
retain university roles — as Hinton has at the
University of Toronto and Ng has at Stanford
University in California — they are usually
only minor, says Bengio. Losing faculty members reduces the number of students that can
be trained, especially at PhD level, adds Abbeel.
Hinton predicts that the shortage in deeplearning experts will be temporary. “The magic
of graduate research in universities is something to be protected, and Google recognizes
that,” he says. Google is currently funding
more than 250 academic research projects and
dozens of PhD fellow­ships.
In supplying industry with talent, universities are fulfilling their natural role, says Michael
Wooldridge, a computer scientist at the University of Oxford,
“There are
UK. And with interso many
est in AI generally
booming, he strugopportunities
gles to see a situation
and so few
in which academia
people to work
is left deserted. The
on them.”
London-based firm
Google DeepMind hired ten researchers from
Oxford in 2014 — but Google gave the university a seven-figure financial contribution, and
formed a research collaboration (see ‘Google
DeepMind’s talent grab’). Many of the poached
staff still hold active teaching positions at the
university — giving students opportunities they
might otherwise never have had.
Bengio is also concerned about the longterm impacts of corporate domination.
Industry researchers are more secretive, he
says. Although scientists in some corporate
labs (such as those at Google and Baidu)
are still publishing papers and code openly
— allowing others to build on their work —
Bengio argues that corporate researchers still
often avoid discussing their work ahead of
publication because they are more likely than
academics to have filed for patents. “That
makes it more difficult to collaborate,” he says.
Some industry insiders are concerned about
transparency, too. In December 2015, SpaceX
founder Elon Musk was among a group of
Silicon Valley investors who launched a nonprofit firm called OpenAI in San Francisco,
California. With $1 billion promised by its
backers, it aims to develop AI for the public
good, sharing its patents and collaborating
freely with other institutions.
Although Google, Facebook and the like
seem committed for the moment to tackling
fundamental questions in AI, Bengio fears
that this might not last. “Business tends to be
pulled to short-term concerns. It’s in the nature
of the beast,” he says. He cites telecommunications firms Bell Labs and AT&T as examples of
companies that had strong research labs, but
eventually lost their talent by putting too much
emphasis on the short-term objective of making money for the company.
Hinton insists that basic research can thrive
in industry. And because of the urgent need
for AI research, some of today’s expansion in
basic research is inevitably taking place at corporate firms, he adds. But academia will still
play a crucial part in AI research, he says. “It’s
the most likely place to get radical new ideas.” ■
.
d
e
v
r
e
s
e
r
s
t
h
g
i
r
l
l
A
.
d
e
t
i
m
i
L
s
r
e
h
s
i
l
b
u
P
n
a
l
l
i
m
c
a
M
6
1
0
2
©
2 8 A P R I L 2 0 1 6 | VO L 5 3 2 | NAT U R E | 4 2 3